Multi-Scale Texture Fusion for Reference-Based Image Super-Resolution: New Dataset and Solution

Download PDF
Open Webpage

Hongyang Zhou, Xiaobin Zhu, Jingyan Qin, Yu Xu, Roberto M. Cesar Jr., Xu-Cheng Yin

Published: 2025, Last Modified: 29 Apr 2026Int. J. Comput. Vis. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Single image super-resolution (SISR) methods often encounter great performance drops on severe degraded low-resolution (LR) images. Recently, reference-based super-resolution (RefSR) methods offer a promising solution by introducing high-quality reference (Ref) images as prior for reconstruction. However, existing RefSR methods often struggle to effectively explore informative textures from Ref images. Additionally, their performance is significantly restricted by the quality of the available training dataset. To address these challenges, we propose an innovative multi-scale texture fusion for reference-based super-resolution via a state-space model, which enables efficient multi-scale feature fusion and long-term dependency modeling for better texture restoration. Specifically, our method mainly consists of a series of texture matching fusion groups (TMFG), which include a multi-scale matching module (MSMM) and a state-space fusion module (SSFM). MSMM can match multi-scale features of Ref images at different stages of the image restoration process to accurately locate key multi-scale similar textures contained in Ref images. SSFM effectively fuses multi-scale textures obtained from Ref images by leveraging stronger modeling relationships based on long-term dependencies of linear complexity. Such a design encourages in-depth exploration of the multi-scale texture correspondence between LR and Ref images, thereby achieving the utilization of textures in Ref and better assisting in image restoration. Notably, we introduce a new large-scale dataset, dubbed DRefSR, designed explicitly for the RefSR task. Our DRefSR offers a wider variety of scenes, more accurately matched image pairs, and a larger volume of samples. With 47,653 image pairs, our dataset substantially exceeds existing datasets (13,761 pairs) in scale. Experiments verify our dataset’s superiority and demonstrate that our method outperforms SOTA methods both quantitatively and qualitatively. DRefSR dataset and code are available at: https://github.com/edbca/SSMTF.